def __init__(self, size):
# Get frame size
x1, y1 = 0, 0
x2, y2 = size
# Resize to fft window
w, h = map(cv2.getOptimalDFTSize, [x2-x1, y2-y1])
x1, y1 = (x1+x2-w)//2, (y1+y2-h)//2
# Store position and size relative to frame
self.pos = x, y = x1+0.5*(w-1), y1+0.5*(h-1)
self.size = w, h
# Create Hanning window (weighting of values from center)
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
# Create output image (centered Gaussian point--for correlation)
g = np.zeros((h, w), np.float32)
g[h//2, w//2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
# Save the desired output image in frequency space
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
# Create sum matrices
self.F = np.zeros_like(self.G)
self.HSum = np.zeros_like(self.G)
self.count = 0
def __init__(self, frame, rect, number):
self.num = number
x1, y1, x2, y2 = rect
w, h = map(cv2.getOptimalDFTSize, [x2-x1, y2-y1])
x1, y1 = (x1+x2-w)//2, (y1+y2-h)//2
self.pos = x, y = x1+0.5*(w-1), y1+0.5*(h-1)
self.size = w, h
img = cv2.getRectSubPix(frame, (w, h), (x, y))
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h//2, w//2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in xrange(128):
a = self.preprocess(MOSSE.rnd_warp(img))
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
self.H2 += cv2.mulSpectrums( A, A, 0, conjB=True)
self.update_kernel()
self.update(frame)
def phase_corr():
single = imageio.imread("images/single-device-mask.jpg", as_gray=True)
grid = imageio.imread("images/raw/01 BF-before.mask.tif")
re = scipy.signal.fftconvolve(single, grid)
cv2.imshow('re', np.log(re / np.max(re)))
cv2.waitKey(0)
cv2.destroyAllWindows()
return
single = pad_fft(single) / 255.
grid = pad_fft(grid) / 255.
height, width = grid.shape
hanw = cv2.createHanningWindow(grid.shape, cv2.CV_64F)
# Windowing and FFT
G_a = np.fft.fft2(single * hanw)
G_b = np.fft.fft2(grid * hanw)
conj_b = np.ma.conjugate(G_b)
R = G_a * conj_b
R /= np.absolute(R)
r = np.fft.fftshift(np.fft.ifft2(R).real)
r = 255. * r / np.max(r)
cv2.imshow('r', np.log(r))
cv2.waitKey(0)
cv2.destroyAllWindows()
def preprocess(self, im):
if self._window is None:
self._window = cv2.createHanningWindow((im.shape[1], im.shape[0]), cv2.CV_64FC1)
# grayscale = np.mean(im, 2)
hanned = im * self._window
image_fft = PhaseCorrelate.preprocessForPhaseCorrelate(hanned)
return image_fft
def PhaseCorrelation(a, b):
height, width = a.shape
#dt = a.dtype # data type
# Windowing
hann_ = cv2.createHanningWindow((height, width), cv2.CV_64F)
#hann = hann_.astype(dt) # convert to correspoinding dtype
rhann = np.sqrt(hann_)
rhann = hann_
# FFT
G_a = np.fft.fft2(a * rhann)
G_b = np.fft.fft2(b * rhann)
conj_b = np.ma.conjugate(G_b)
R = G_a * conj_b
R /= np.absolute(R)
r = np.fft.fftshift(np.fft.ifft2(R).real)
# Get result and Interpolation
DY, DX = np.unravel_index(r.argmax(), r.shape)
# Subpixel Accuracy
boxsize = 5
box = r[DY - int((boxsize - 1) / 2):DY + int((boxsize - 1) / 2) + 1,
DX - int((boxsize - 1) / 2):DX + int((boxsize - 1) / 2) +
1] # x times x box
#TY,TX= CenterOfGravity(box)
TY, TX = WeightedCOG(box)
sDY = TY + DY
sDX = TX + DX
# Show the result
# print('DX=',width/2-sDX,'DY=',height/2-sDY)
# print('CorrelationVal=',r[DY,DX])
plt.imshow(r, vmin=r.min(), vmax=r.max())
return [width / 2 - sDX, height / 2 - sDY], r[DY, DX]
def __init__(self, frame, rect, index, prev_id=0, reInit=False):
self.stationary_for_frames = 0
x1, y1, x2, y2 = rect
self.index = index
if not reInit:
self.id_ = MOSSE.MosseCounter
MOSSE.MosseCounter = MOSSE.MosseCounter + 1
else:
self.id_ = prev_id
w, h = map(cv2.getOptimalDFTSize, [x2 - x1, y2 - y1])
x1, y1 = (x1 + x2 - w) // 2, (y1 + y2 - h) // 2
self.pos = x, y = x1 + 0.5 * (w - 1), y1 + 0.5 * (h - 1)
self.size = w, h
img = cv2.getRectSubPix(frame, (w, h), (x, y))
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h // 2, w // 2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in xrange(128):
a = self.preprocess(rnd_warp(img))
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
self.H2 += cv2.mulSpectrums(A, A, 0, conjB=True)
self.update_kernel()
self.update(frame)
def __init__(self, frame, rect):
org = frame.copy()
if len(frame.shape) == 3:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
x1, y1, x2, y2 = rect
w, h = map(cv2.getOptimalDFTSize, [x2-x1, y2-y1])
x1, y1 = (x1+x2-w)//2, (y1+y2 - h)//2
self.pos = x,y = x1+0.5*(w-1), y1 + 0.5*(h-1)
self.size = w, h
img = cv2.getRectSubPix(frame, (w,h), (x,y))
self.win = cv2.createHanningWindow((w,h), cv2.CV_32F)
#expected response
g = np.zeros((h,w), np.float32)
g[h//2, w//2] = 1
g = cv2.GaussianBlur(g, (-1,-1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags = cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in range(128):
f = self.preprocess(self.rnd_warp(img), self.win)
F = cv2.dft(f, flags = cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, F, 0, conjB = True)
self.H2 += cv2.mulSpectrums( F, F, 0, conjB = True)
self.update_kernel() #update kernel
self.update(frame) #
def shift_stabilization(img1, img2, rows, cols, print_result=None):
"""
Perform shift stabilization on two images using phase correlation with hanning window
:param img1 source image
:param img2: target image
:param rows: rows of result image
:param cols: columns of result image
:param print_result: gathered information during stabilization
:return:
result_image: stabilized (shifted) image
print_result: collected information during shift stabilization
"""
img1_gray = ImageProcess.to_gray(img1)
img2_gray = ImageProcess.to_gray(img2)
hanning = cv2.createHanningWindow((cols, rows), cv2.CV_32F)
(cx, cy), _ = cv2.phaseCorrelate(np.float32(img2_gray),
np.float32(img1_gray))
# (cx, cy) = (round(cx, 2), round(cy, 2))
M = np.float32([[1, 0, cx], [0, 1, cy]])
print_result['x'] = cx
print_result['y'] = cy
t_form = transform.EuclideanTransform(translation=(cx, cy))
result_image = transform.warp(img2, t_form)
return img_as_ubyte(result_image), print_result
def onrect(self, rect):
#print "onrect enterd:",rect
self.lk_ready = True
frame_gray = cv2.cvtColor(self.frame, cv2.COLOR_BGR2GRAY)
frame_gray = cv2.equalizeHist(frame_gray)
self.old_gray = frame_gray.copy()
self.patch = self.old_gray[rect[1]:rect[3],rect[0]:rect[2]]
self.w = rect[2]-rect[0]
self.h = rect[3]-rect[1]
#cv2.imshow('patch',self.patch)
#cv2.waitKey(10)
mask = cv2.createHanningWindow((self.w,self.h),cv2.CV_32F)
self.patch = np.uint8(self.patch*mask)
self.p0 = getUniformPoints(self.old_gray,rect[:2],rect[2:],100)
# self.p0 = cv2.goodFeaturesToTrack(self.patch, mask = None, **feature_params)
if len(self.p0)==0:
self.lk_ready = False
print "len p0:",len(self.p0)
## else:
## for i in xrange(len(self.p0)):
## self.p0[i][0][0] = self.p0[i][0][0] + rect[0]
## self.p0[i][0][1] = self.p0[i][0][1] + rect[1]
tracker = MOSSE(frame_gray, rect)
self.trackers.append(tracker)
def init(self, frame, rect):
if frame.ndim == 3:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
rect = rect.astype(int)
rect[2:] += rect[:2]
x1, y1, x2, y2 = rect
w, h = map(cv2.getOptimalDFTSize, [x2 - x1, y2 - y1])
x1, y1 = (x1 + x2 - w) // 2, (y1 + y2 - h) // 2
self.t_center = x, y = x1 + 0.5 * (w - 1), y1 + 0.5 * (h - 1)
self.t_sz = w, h
img = cv2.getRectSubPix(frame, (w, h), (x, y))
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h // 2, w // 2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.A = np.zeros_like(self.G)
self.B = np.zeros_like(self.G)
for _i in range(128):
patch = self._preprocess(self._random_warp(img))
F = cv2.dft(patch, flags=cv2.DFT_COMPLEX_OUTPUT)
self.A += cv2.mulSpectrums(self.G, F, 0, conjB=True)
self.B += cv2.mulSpectrums(F, F, 0, conjB=True)
self._update_kernel()
self.update(frame)
def __init__(self, frame, rect):
x1, y1, x2, y2 = rect
w, h = map(cv2.getOptimalDFTSize, [x2 - x1, y2 - y1])
x1, y1 = (x1 + x2 - w) // 2, (y1 + y2 - h) // 2
self.pos = x, y = x1 + 0.5 * (w - 1), y1 + 0.5 * (h - 1)
self.size = w, h
#self.good = 0
img = cv2.getRectSubPix(frame, (w, h), (x, y))
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h // 2, w // 2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in xrange(128):
a = self.preprocess(rnd_warp(img))
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
self.H2 += cv2.mulSpectrums(A, A, 0, conjB=True)
self.update_kernel()
self.update(frame)
def pre_process(img):
log_img = np.log(img + 1)
normalize_img = (log_img - np.mean(log_img.flatten())) / (
np.std(log_img.flatten() + 1e-5))
cosine_window = cv2.createHanningWindow(img.shape[:2], cv2.CV_32F)
output = normalize_img * cosine_window.T
return output
def __init__(self, frame, rect):
x1, y1, x2, y2 = rect # Extract coordinates of the rectangle to be tracked
w, h = map(cv2.getOptimalDFTSize, [x2 - x1, y2 - y1])
x1, y1 = (x1 + x2 - w) // 2, (y1 + y2 - h) // 2
# 0.5 * x \in [w, h] (-1) = the centre point of the region
self.pos = x, y = x1 + 0.5 * (w - 1), y1 + 0.5 * (h - 1)
self.size = w, h
img = cv2.getRectSubPix(frame, (w, h), (x, y))
# Hanning Window
# http://en.wikipedia.org/wiki/Window_function
# http://en.wikipedia.org/wiki/Window_function#Hann_.28Hanning.29_window
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h // 2, w // 2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, None, cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in xrange(128):
a = self.preprocess(rnd_warp(img))
A = cv2.dft(a, None, cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
self.H2 += cv2.mulSpectrums(A, A, 0, conjB=True)
self.update_kernel()
self.update(frame)
self.id = id(self)
def make_image():
image = np.random.rand(20, 70, 3)
window = cv2.createHanningWindow((image.shape[1], image.shape[0]), cv2.CV_64FC1)
grayscale = np.mean(image, 2)
hanned = grayscale * window
image_fft = PhaseCorrelate.preprocessForPhaseCorrelate(hanned)
return image, image_fft
def __reduce_image_noise_and_details(self):
# apply Gaussian Blur on the original image, to reduce noise and the amount of details, for
# faster processing
self.g = np.zeros((self.size[::-1]), np.float32)
self.g[self.size[1] // 2, self.size[0] // 2] = 1
self.g = cv2.GaussianBlur(self.g, (-1, -1), 2.0)
self.g /= self.g.max()
self.hann_window = cv2.createHanningWindow(self.size, cv2.CV_32F)
def getCylHanning(size, axis=0, ptype=cv2.CV_64F):
tmp = cv2.createHanningWindow(size[::-1], ptype)
if axis == 0:
tmp = tmp[size[1] / 2, :]
ret = np.tile(tmp, (size[0], 1))
else:
tmp = tmp[:, size[0] / 2]
ret = np.transpose(np.tile(tmp, (size[1], 1)))
# print ret.shape
return ret
def getCylHanning(size, axis=0, ptype=cv2.CV_64F):
tmp=cv2.createHanningWindow(size[::-1], ptype)
if axis==0:
tmp=tmp[size[1]/2,:]
ret=np.tile(tmp,(size[0],1))
else:
tmp=tmp[:, size[0]/2]
ret=np.transpose(np.tile(tmp,(size[1],1)))
# print ret.shape
return ret
def pre_process(img):
# get the size of the img...
height, width = img.shape
img = np.log(np.float32(img) + 1.0)
img = (img - img.mean()) / (img.std() + 1e-5)
# use the hanning window...
# window = window_func_2d(height, width)
window = cv2.createHanningWindow((width, height), cv2.CV_32F)
img = img * window
return img
def initialise(self, gray_image, bounding_box):
'''
Receives a grayscale frame and a bounding box and operates over it
return True if OK
Bounding box format: ((x0, y0),(x1, y1))
This uses the Numpy FFT since it already represents the numbers in
complex representation and allows to do Spectrum Multiplciation and
Vision in a straight-forward fashion
'''
self.last_frame = gray_image
p1, w, h = self.extractBoundingBox(bounding_box)
p1, p2 = bounding_box
cols = p2[0] - p1[0]
rows = p2[1] - p1[1]
self.size = (w, h)
x1 = int(round((2 * p1[0] + cols - w) / 2))
y1 = int(round((2 * p1[1] + rows - h) / 2))
self.center = (x1 + (w / 2), y1 + (h / 2))
# Get inputs - window from BBox and the HanningWindow for gaussianity
window = cv.getRectSubPix(gray_image, self.size, self.center)
self.hanWin = cv.createHanningWindow(self.size, cv.CV_32F)
# Create goal and its FFT
g = np.zeros((h, w), np.float32)
g[int(h / 2)][int(w / 2)] = 1.
g = cv.GaussianBlur(g, (-1, -1), 2.0)
maxVal = cv.minMaxLoc(g)[1]
maxVal = 1. / maxVal
g = g * maxVal
self.G = np.fft.fft2(g)
# Initialise the filter and other elements
self.A = np.zeros(self.G.shape, dtype=np.complex128)
self.B = np.zeros(self.G.shape, dtype=np.complex128)
# Train the filter with a random warping
for i in range(8):
warped = randWarp(window)
self.f = preprocess(warped, self.hanWin)
F = np.fft.fft2(self.f)
A_i = self.G * np.conjugate(F)
B_i = F * np.conjugate(F)
self.A += A_i
self.B += B_i
mask_b = np.ma.array(self.B, mask=(self.B == 0))
self.H = self.A / mask_b
self.H = self.H.filled(0.)
return True
def create_window(reference, plot=False):
height, width = reference.shape
hanning_window = cv2.createHanningWindow((width, height), cv2.CV_64F)
#hanning_window = hanning_window[height//2, :]
#hanning_window = np.outer(hanning_window, hanning_window)
if plot:
plt.contourf(hanning_window)
plt.show()
return hanning_window
def phaseCorr(img1, img2, isUseHann=True, isDebug=False):
assert(img1.shape==img2.shape)
if isUseHann:
hw=cv2.createHanningWindow(img1.shape[::-1], cv2.CV_64F)
img1=img1*hw
img2=img2*hw
if isDebug:
cv2.imshow("debug-img12", np.concatenate((nrmMat(img1), nrmMat(img2))))
fft1=np.fft.fft2(img1)
fft2=np.fft.fft2(img2)
CC=np.fft.fftshift(np.fft.ifft2( (fft1*fft2.conj())/(fft2*fft2.conj()) ).real)
minVal, maxVal, minLoc, maxLoc=cv2.minMaxLoc(CC)
dxy=(maxLoc - np.array(img1.shape[::-1])/2)
return (CC, dxy, maxVal)
def phaseCorr(img1, img2, isUseHann=True, isDebug=False):
assert (img1.shape == img2.shape)
if isUseHann:
hw = cv2.createHanningWindow(img1.shape[::-1], cv2.CV_64F)
img1 = img1 * hw
img2 = img2 * hw
if isDebug:
cv2.imshow("debug-img12", np.concatenate((nrmMat(img1), nrmMat(img2))))
fft1 = np.fft.fft2(img1)
fft2 = np.fft.fft2(img2)
CC = np.fft.fftshift(
np.fft.ifft2((fft1 * fft2.conj()) / (fft2 * fft2.conj())).real)
minVal, maxVal, minLoc, maxLoc = cv2.minMaxLoc(CC)
dxy = (maxLoc - np.array(img1.shape[::-1]) / 2)
return (CC, dxy, maxVal)
def ripoc(srcImg1, srcImg2, logpolarMagnitude=40, plotFig=False):
grayImg1 = np.float64(cv2.cvtColor(srcImg1, cv2.COLOR_BGRA2GRAY)) / 255.0
grayImg2 = np.float64(cv2.cvtColor(srcImg2, cv2.COLOR_BGRA2GRAY)) / 255.0
center = (grayImg1.shape[0] / 2.0, grayImg1.shape[1] / 2.0)
# lpImg1 = cv2.logPolar(grayImg1, center, logpolarMagnitude, cv2.INTER_CUBIC + cv2.WARP_FILL_OUTLIERS)
lpImg1 = _logPolar(grayImg1, center, logpolarMagnitude)
center = (grayImg2.shape[0] / 2.0, grayImg2.shape[1] / 2.0)
# lpImg2 = cv2.logPolar(grayImg2, center, logpolarMagnitude, cv2.INTER_CUBIC + cv2.WARP_FILL_OUTLIERS)
lpImg2 = _logPolar(grayImg2, center, logpolarMagnitude)
if plotFig:
figLP, (figLPLeft, figLPRight) = plt.subplots(ncols=2)
figLPLeft.imshow(lpImg1, cmap="gray")
figLPRight.imshow(lpImg2, cmap="gray")
plt.show(figLP)
hann = cv2.createHanningWindow(lpImg1.shape, cv2.CV_64F)
ret, resp = cv2.phaseCorrelate(lpImg1, lpImg2, hann)
# angle = -((ret[1] + 0.5) * 360.0 / float(lpImg1.shape[0]) )
# scale = np.exp(ret[0] / logpolarMagnitude)
angle = ret[1] * 360.0 / float(lpImg1.shape[0])
scale = np.exp(ret[0] / logpolarMagnitude)
# print("ret = " + str(ret) + "\nresponse = " + str(resp))
# print("angle = " + str(angle) + "\nscale = " + str(scale))
matTrans = cv2.getRotationMatrix2D(center, angle, scale)
imgTrans = cv2.warpAffine(srcImg1,
matTrans,
srcImg1.shape[0:2],
flags=cv2.INTER_LINEAR)
if plotFig:
figDst, (figDstLeft, figDstRight) = plt.subplots(ncols=2)
figDstLeft.imshow(imgTrans)
figDstRight.imshow(srcImg2)
plt.show(figDst)
subtractImg = cv2.subtract(cv2.cvtColor(imgTrans, cv2.COLOR_BGRA2RGB),
cv2.cvtColor(srcImg2, cv2.COLOR_BGRA2RGB))
plt.imshow(subtractImg)
plt.show()
grayTrans = np.float64(cv2.cvtColor(imgTrans, cv2.COLOR_BGRA2GRAY)) / 255.0
ret, resp = cv2.phaseCorrelate(grayTrans, grayImg2, hann)
return angle, scale, ret, resp
def train_H(self, train_data):
h, w = train_data[0].shape
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h // 2, w // 2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for elem in train_data:
a = self.preprocess(elem)
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(
self.G, A, 0, conjB=True) # conjB=True 表示在做乘法之前取第二个输入数组的共轭.
self.H2 += cv2.mulSpectrums(A, A, 0, conjB=True)
self.update_kernel()
def phaseCorr(img1, img2, isUseHann=True, isDebug=False, isFloatAcc=True, RadPQ=3):
assert(img1.shape==img2.shape)
if isUseHann:
hw=cv2.createHanningWindow(img1.shape[::-1], cv2.CV_64F)
img1=img1*hw
img2=img2*hw
if isDebug:
cv2.imshow("debug-img12", np.concatenate((nrmMat(img1), nrmMat(img2))))
fft1=np.fft.fft2(img1)
fft2=np.fft.fft2(img2)
CC=np.fft.fftshift(np.fft.ifft2( (fft1*fft2.conj())/(fft2*fft2.conj()) ).real)
minVal, maxVal, minLoc, maxLoc=cv2.minMaxLoc(CC)
_,_,pq=calcPeakQ(CC,maxLoc,RadPQ)
if isFloatAcc:
maxLoc=calcWeightedCenter(CC,maxLoc,RadPQ)
dxy=(maxLoc - np.array(img1.shape[::-1])/2) #TODO: Check this point !!!
return (CC, dxy, maxVal,pq)
def phaseCorr(img1, img2, isUseHann=True, isDebug=False, isFloatAcc=True, RadPQ=3):
assert(img1.shape==img2.shape)
if isUseHann:
hw=cv2.createHanningWindow(img1.shape[::-1], cv2.CV_64F)
img1=img1*hw
img2=img2*hw
if isDebug:
cv2.imshow("debug-img12", np.concatenate((nrmMat(img1), nrmMat(img2))))
fft1=np.fft.fft2(img1)
fft2=np.fft.fft2(img2)
CC=np.fft.fftshift(np.fft.ifft2( (fft1*fft2.conj())/(fft2*fft2.conj()) ).real)
minVal, maxVal, minLoc, maxLoc=cv2.minMaxLoc(CC)
_,_,pq=calcPeakQ(CC,maxLoc,RadPQ)
if isFloatAcc:
maxLoc=calcWeightedCenter(CC,maxLoc,RadPQ)
dxy=(maxLoc - np.array(img1.shape[::-1])/2) #TODO: Check this point !!!
return (CC, dxy, maxVal,pq)
def __init__(self, frame, rect):
x1, y1, x2, y2 = rect
self.frameWidth = frame.shape[1]
self.frameHeight = frame.shape[0]
self.firstframe = frame[int(y1):int(y2), int(x1):int(x2)]
cv2.imshow('first frame', self.firstframe)
fout.write('bounding box initial position : ')
fout.write(
str(x1) + ',' + str(y1) + ',' + str(x2 - x1) + ',' + str(y2 - y1) +
'\n')
fout.write('tracking object from here \n')
w, h = map(cv2.getOptimalDFTSize, [x2 - x1, y2 - y1])
x1, y1 = (x1 + x2 - w) // 2, (y1 + y2 - h) // 2
self.pos = x, y = x1 + 0.5 * (w - 1), y1 + 0.5 * (h - 1)
self.size = w, h
img = cv2.getRectSubPix(frame, (w, h), (x, y))
self.PF_number = n_PF
self.prev_PF_count = n_PF
self.f0 = np.array([])
self.f = np.array([])
self.pt = np.ones((n_PF, 2), int) * self.pos
self.last_img_PF = np.array([])
self.last_resp_PF = np.array([])
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h // 2, w // 2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in xrange(128):
a = self.preprocess(rnd_warp(img))
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
self.H2 += cv2.mulSpectrums(A, A, 0, conjB=True)
self.update_kernel()
def __init__(self, frame, rect):
# Get frame size
x1, y1, x2, y2 = rect
# Resize to fft window
w, h = map(cv2.getOptimalDFTSize, [x2-x1, y2-y1])
x1, y1 = (x1+x2-w)//2, (y1+y2-h)//2
# Store position and size relative to frame
self.pos = x, y = x1+0.5*(w-1), y1+0.5*(h-1)
self.size = w, h
# Crop template image from frame
img = cv2.getRectSubPix(frame, (w, h), (x, y))
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# Create Hanning window (weighting of values from center)
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
# Create output image (centered Gaussian point--for correlation)
g = np.zeros((h, w), np.float32)
g[h//2, w//2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
# Save the desired output image in frequency space
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
# Create transformed variants of input
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in xrange(128):
# Preprocess, get DFT
a = self.preprocess(rnd_warp(img))
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True) # Sum of G x F*
self.H2 += cv2.mulSpectrums( A, A, 0, conjB=True) # Sum of F x F*
# Update filter
self.update_kernel()
self.update(frame)
def calcPhaseCorrShift(fimg1, fimg2):
frm1=cv2.imread(fimg1, 0)
frm2=cv2.imread(fimg2, 0)
frm1=cv2.resize(frm1.astype(np.float).copy(), (int(frm1.shape[1]/kdif), int(frm1.shape[0]/kdif)))
frm2=cv2.resize(frm2.astype(np.float).copy(), (int(frm2.shape[1]/kdif), int(frm2.shape[0]/kdif)))
frm1_nrm=cv2.normalize(frm1.copy(), None, 0,255, cv2.NORM_MINMAX, cv2.CV_8U)
frm2_nrm=cv2.normalize(frm2.copy(), None, 0,255, cv2.NORM_MINMAX, cv2.CV_8U)
hann=cv2.createHanningWindow((frm1.shape[1], frm1.shape[0]), cv2.CV_64F)
#
shift = cv2.phaseCorrelate(frm1, frm2, hann)
dxy=shift[0]
frm2_nrm_shift=np.roll(frm2_nrm, int(math.floor(-dxy[0])), 1)
frm2_nrm_shift=np.roll(frm2_nrm_shift, int(math.floor(-dxy[1])), 0)
tmp=np.zeros((frm1.shape[0], frm1.shape[1], 3), np.uint8)
tmp[:,:,2]=frm1_nrm
tmp[:,:,1]=frm2_nrm_shift
tmp[:,:,0]=0
# tmp[:,:,1]=frm2_nrm
cv2.imshow("frame #2 shift", cv2.resize(tmp, (tmp.shape[1]/1, tmp.shape[0]/1)))
return dxy
def __init__(self, frame, rect):
x1, y1, x2, y2 = rect
self.frameWidth = frame.shape[1]
self.frameHeight = frame.shape[0]
self.firstframe = frame[int(y1):int(y2),int(x1):int(x2)]
cv2.imshow('first frame',self.firstframe)
fout.write('bounding box initial position : ')
fout.write(str(x1)+','+ str(y1)+','+ str(x2-x1)+','+ str(y2-y1)+'\n')
fout.write('tracking object from here \n')
w, h = map(cv2.getOptimalDFTSize, [x2-x1, y2-y1])
x1, y1 = (x1+x2-w)//2, (y1+y2-h)//2
self.pos = x, y = x1+0.5*(w-1), y1+0.5*(h-1)
self.size = w, h
img = cv2.getRectSubPix(frame, (w, h), (x, y))
self.PF_number = n_PF
self.prev_PF_count = n_PF
self.f0 = np.array([])
self.f = np.array([])
self.pt = np.ones((n_PF, 2), int) * self.pos
self.last_img_PF = np.array([])
self.last_resp_PF = np.array([])
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h//2, w//2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
for i in xrange(128):
a = self.preprocess(rnd_warp(img))
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
self.H2 += cv2.mulSpectrums( A, A, 0, conjB=True)
self.update_kernel()
def register(fn1, fn2):
im1 = cv2.imread(fn1)
im2 = cv2.imread(fn2)
# Create a Hanning window.
print(im1.shape)
im1_flt = tofloat(im1)
im2_flt = tofloat(im2)
hann = cv2.createHanningWindow(im1.shape[:2], cv2.CV_64F)
# Find the shift using.
shift = cv2.phaseCorrelate(im1_flt, im2_flt)
print('shift: ', shift)
radius = math.sqrt(shift.x * shift.x + shift.y*shift.y)
if radius > 5:
center = (im1.cols >> 1, im1.rows >> 1)
cv2.circle(im1, center, int(radius), (0,255,0), 3)
cv2.imshow('result', im1)
cv2.waitKey(0)
def __init__(self, frame, rect, trackNo, frameNo):
x1, y1, x2, y2 = rect
self.startpos = (x1, y1)
self.startFrame = frameNo
w, h = map(cv2.getOptimalDFTSize, [x2 - x1, y2 - y1])
x1, y1 = (x1 + x2 - w) // 2, (y1 + y2 - h) // 2
self.pos = x, y = x1 + 0.5 * (w - 1), y1 + 0.5 * (h - 1)
#self.lastPos = self.pos
self.size = w, h
self.objectMoved = False
self.objectLeft = False
self.stationaryCount = 0
#self.isOverLap = False
#print(frame,w,h,x,y)
self.trackNo = trackNo
img = cv2.getRectSubPix(frame, (w, h), (x, y))
self.obj = None
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
g = np.zeros((h, w), np.float32)
g[h // 2, w // 2] = 1
g = cv2.GaussianBlur(g, (-1, -1), 2.0)
g /= g.max()
self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 = np.zeros_like(self.G)
self.H2 = np.zeros_like(self.G)
# print('in constructor - img = {}, img.shape = {}'.format(img, img.shape))
for i in range(128):
a = self.preprocess(Apply_affine(img))
A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
self.H2 += cv2.mulSpectrums(A, A, 0, conjB=True)
self.update_Filter()
self.update(frame)
def __init__(self, frame, rect): #given a initial frame, and ROI region (top-left,down-right)
if len(frame.shape) == 3:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
x1, y1, x2, y2 = rect
w_, h_ = x2-x1, y2-y1
self.factor = 2 # region is factor * roi
x1, y1 = (x1+x2- w_)//2, (y1+y2 - h_ )//2
self.pos = x,y = x1+0.5*(w_-1), y1 + 0.5*(h_-1) #center
w, h = map(cv2.getOptimalDFTSize, [int(self.factor*w_), int(self.factor*h_)]) #get optimal DFT size
self.size = w, h
img = cv2.getRectSubPix(frame, self.size, self.pos) #will auto padding if size exceeds boundaries
self.win = cv2.createHanningWindow(self.size, cv2.CV_32F)
self.X = self.preprocess(img, self.win) #first patch
self.Y = self.target(w/self.factor, h/self.factor, self.factor) #expected response, gaussian distri.
self.sigma = 0.2
self.lamb = 0.01
self.dgk = self.gaussian_kernel
self.alphaf = self.training(self.X, self.Y, self.sigma, self.lamb)
def test(self, frame):
(x,y), (w,h) = self.pos, self.size
win = cv2.createHanningWindow((w,h), cv2.CV_32F)
img1 = self.preprocess(self.last_img, win)
F1 = cv2.dft(img1, flags = cv2.DFT_COMPLEX_OUTPUT)
G1 = cv2.mulSpectrums(F1, self.H, 0, conjB = True)
resp1 = cv2.idft(G1, flags = cv2.DFT_SCALE|cv2.DFT_REAL_OUTPUT)
img = cv2.getRectSubPix(frame, (w, h), (x,y))
img2 = self.preprocess(img, win)
F2 = cv2.dft(img2, flags = cv2.DFT_COMPLEX_OUTPUT)
G2 = cv2.mulSpectrums(F2, self.H, 0, conjB = True)
resp2 = cv2.idft(G2, flags = cv2.DFT_SCALE|cv2.DFT_REAL_OUTPUT)
merge1 = np.hstack((img1 , resp1))
merge2 = np.hstack((img2, resp2))
merge = np.vstack((merge1, merge2))
name = 'comparison--last vs cur.'
cv2.namedWindow(name, cv2.WINDOW_NORMAL)
cv2.resizeWindow(name, 600,400)
cv2.imshow(name, merge)
def match(self):
height, width = self.ref.shape
self.hanw = cv2.createHanningWindow((height, width), cv2.CV_64F)
# Windowing and FFT
G_a = np.fft.fft2(self.ref * self.hanw)
G_b = np.fft.fft2(self.cmp * self.hanw)
# 1.1: Frequency Whitening
self.LA = np.fft.fftshift(np.log(np.absolute(G_a) + 1))
self.LB = np.fft.fftshift(np.log(np.absolute(G_b) + 1))
# 1.2: Log polar Transformation
cx = self.center[1]
cy = self.center[0]
self.Mag = width / (math.log(width) - math.log(2) * 0.5)
self.LPA = cv2.logPolar(self.LA, (cy, cx),
self.Mag,
flags=cv2.INTER_LINEAR +
cv2.WARP_FILL_OUTLIERS)
self.LPB = cv2.logPolar(self.LB, (cy, cx),
self.Mag,
flags=cv2.INTER_LINEAR +
cv2.WARP_FILL_OUTLIERS)
# 1.3:filtering
LPmin = math.floor(self.Mag *
math.log(self.alpha * width / 2.0 / math.pi))
LPmax = min(width,
math.floor(self.Mag * math.log(width * self.beta / 2)))
assert LPmax > LPmin, 'Invalid condition!\n Enlarge lpmax tuning parameter or lpmin_tuning parameter'
Tile = np.repeat([0.0, 1.0, 0.0],
[LPmin - 1, LPmax - LPmin + 1, width - LPmax])
self.Mask = np.tile(Tile, [height, 1])
self.LPA_filt = self.LPA * self.Mask
self.LPB_filt = self.LPB * self.Mask
# 1.4: Phase Correlate to Get Rotation and Scaling
Diff, peak, self.r_rotatescale = self.PhaseCorrelation(
self.LPA_filt, self.LPB_filt)
theta1 = 2 * math.pi * Diff[1] / height
# deg
theta2 = theta1 + math.pi
# deg theta ambiguity
invscale = math.exp(Diff[0] / self.Mag)
# 2.1: Correct rotation and scaling
b1 = self.Warp_4dof(self.cmp, [0, 0, theta1, invscale])
b2 = self.Warp_4dof(self.cmp, [0, 0, theta2, invscale])
# 2.2 : Translation estimation
diff1, peak1, self.r1 = self.PhaseCorrelation(
self.ref, b1) #diff1, peak1 = PhaseCorrelation(a,b1)
diff2, peak2, self.r2 = self.PhaseCorrelation(
self.ref, b2) #diff2, peak2 = PhaseCorrelation(a,b2)
# Use cv2.phaseCorrelate(a,b1) because it is much faster
# 2.3: Compare peaks and choose true rotational error
if peak1 > peak2:
Trans = diff1
peak = peak1
theta = -theta1
else:
Trans = diff2
peak = peak2
theta = -theta2
if theta > math.pi:
theta -= math.pi * 2
elif theta < -math.pi:
theta += math.pi * 2
# Results
self.param = [Trans[0], Trans[1], theta, 1 / invscale]
self.peak = peak
self.perspective = self.poc2warp(self.center, self.param)
self.affine = self.perspective[0:2, :]
def _process_movie(self, **kwargs):
"""
Function for analyzing a full film or video. This is where the magic happens when we're making analyses. Function will exit if neither a movie path nor a data path are supplied. This function is not intended to be called directly. Normally, call one of the two analyze_ functions instead, which will call this function.
"""
ap = self._check_pcorr_params(kwargs)
verbose = ap['verbose']
if not HAVE_CV:
print "WARNING: You must install OpenCV in order to analyze or view!"
return
if (self.movie_path is None) and (self.data_path is None) and ap['mode'] is 'analysis':
print "ERROR: Must supply both a movie and a data path for analysis!"
return
have_mov = os.path.exists(self.movie_path)
have_data = os.path.exists(self.data_path)
if (have_mov is False) and (have_data is False):
print "Both movie file and data file are missing! Please supply at least one."
return None
print ap
if have_mov:
self.capture = cv2.VideoCapture(self.movie_path)
frame_width = int(self.capture.get(cv.CV_CAP_PROP_FRAME_WIDTH))
frame_height = int(self.capture.get(cv.CV_CAP_PROP_FRAME_HEIGHT))
else:
frame_width = 640
frame_height = int(frame_width / self._read_json_value('aspect'))
fps = ap['fps']
grid_x_divs = ap['grid_divs_x']
grid_y_divs = ap['grid_divs_y']
frame_size = (frame_width, frame_height)
grid_width = int(frame_width/grid_x_divs)
grid_height = int(frame_height/grid_y_divs)
grid_size = (grid_width, grid_height)
centers_x = range((frame_width/16),frame_width,(frame_width/8))
centers_y = range((frame_height/16),frame_height,(frame_height/8))
if verbose:
print fps, ' | ', frame_size, ' | ', grid_size
# container for prev. frame's grayscale subframes
prev_sub_grays = []
if ap['offset'] > 0:
offset_secs = ap['offset']
else:
offset_secs = 0
print "%%% ", ap['duration']
if ap['duration'] > 0:
dur_secs = ap['duration']
elif ap['duration'] < 0:
ap['duration'] = self.determine_movie_length()
else:
print "Duration cannot be 0."
return
dur_secs = ap['duration']
stride_frames = ap['stride']
stride_hop = stride_frames - 1
print "1. ", ap['duration']
print "2. ", dur_secs
# check offset first, then compress duration, if needed
offset_secs = min(max(offset_secs, 0), ap['duration'])
dur_secs = min(max(dur_secs, 0), (ap['duration'] - offset_secs))
offset_strides = int(offset_secs * (fps / stride_frames))
dur_strides = int(dur_secs * (fps / stride_frames))
offset_frames = offset_strides * stride_frames
dur_frames = dur_strides * stride_frames
if verbose:
print '%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%'
print 'DUR TOTAL: ', dur_total_secs
print "OFFSET (SECONDS): ", offset_secs
print "OFFSET (STRIDES): ", offset_strides
print "OFFSET (FRAMES): ", offset_frames
print "DUR (SECONDS): ", dur_secs
print 'DUR (STRIDES): ', dur_strides
print 'DUR (FRAMES): ', dur_frames
print 'XDIVS: ', grid_x_divs
print 'YDIVS: ', grid_y_divs
print "FPS: ", fps
print "stride_frames: ", stride_frames
# set up memmap
if ap['mode'] == 'playback' and ap['display'] == True:
fp = np.memmap(self.data_path, dtype='float32', mode='r', shape=((offset_strides + dur_strides),(64+1),2))
else:
fp = np.memmap(self.data_path, dtype='float32', mode='w+', shape=(dur_strides,(64+1),2))
if ap['display']:
cv.NamedWindow('Image', cv.CV_WINDOW_AUTOSIZE)
cv.ResizeWindow('Image', frame_width, frame_height)
self.frame_idx = offset_frames
end_frame = offset_frames + dur_frames
if have_mov:
self.capture.set(cv.CV_CAP_PROP_POS_FRAMES, offset_frames)
ret, frame = self.capture.read()
if frame is None:
print 'Frame error! Exiting...'
return # no image captured... end the processing
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
prev_frame_gray = np.float32(frame_gray[:])
fhann = cv2.createHanningWindow((frame_width,frame_height), cv2.CV_32FC1)
ghann = cv2.createHanningWindow((grid_width,grid_height), cv2.CV_32FC1)
for row in range(grid_y_divs):
for col in range(grid_x_divs):
prev_sub_grays += [np.float32(frame_gray[(row*grid_height):((row+1)*grid_height), (col*grid_width):((col+1)*grid_width)])]
else:
frame = np.empty((frame_width, frame_height), np.uint8)
print frame.shape
self.frame_idx += 1
print "<<< ", frame.mean()
# cv.ShowImage('Image', cv.fromarray(frame))
while self.frame_idx < end_frame:
if (self.frame_idx % 1000) == 0: print 'fr. idx: ', self.frame_idx, ' (', self.frame_idx / float(end_frame), ' | ', end_frame, ')'
if have_mov:
# grab next frame
ret, frame = self.capture.read()
if frame is None:
print 'Frame error! Exiting...'
break # no image captured... end the processing
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
else:
frame[:] = 0
# display stage (full)
if ap['mode'] == 'playback' and ap['display']:
fret = fp[self.frame_idx][64]
# print fret
elif have_mov:
fret, fres = cv2.phaseCorrelateRes(prev_frame_gray, np.float32(frame_gray[:]), fhann)
print fret
if abs(fres) > 0.01:
fp[self.frame_idx][64] = [(fret[0]/frame_width),(fret[1]/frame_height)]
else:
fp[self.frame_idx][64] = [0,0]
else:
return
# display stage (gridded)
for row in range(grid_y_divs):
for col in range(grid_x_divs):
if ap['mode'] == 'playback' and ap['display']:
cell = ((row*8)+col)
gret = fp[self.frame_idx][cell]
elif have_mov:
sub_gray = np.float32(frame_gray[(row*grid_height):((row+1)*grid_height), (col*grid_width):((col+1)*grid_width)][:])
gret, gres = cv2.phaseCorrelateRes(prev_sub_grays[(row*grid_x_divs)+col], sub_gray, ghann)
prev_sub_grays[(row*grid_x_divs)+col] = sub_gray
if abs(gres) > 0.7: # WAS 0.01!!!!
fp[self.frame_idx][(row*grid_x_divs)+col] = [(gret[0]/grid_width),(gret[1]/grid_height)]
else:
fp[self.frame_idx][(row*grid_x_divs)+col] = [0,0]
else:
return
if ap['display'] and (gret[0] != 0 and gret[1] != 0):
print gret
print grid_size
print centers_x[col]
print centers_y[row]
xval = int(min((gret[0]*1000), grid_size[0])+centers_x[col])
yval = int(min((gret[1]*1000), grid_size[1])+centers_y[row])
# print ((centers_x[i], centers_y[j], xval, yval), False, (0,255,255))
print (centers_x[col], centers_y[row])
print (xval, yval)
cv2.line(frame, (centers_x[col], centers_y[row]), (xval, yval), (255,255,255))
#### SHOW
if ap['display'] and have_mov:
print "<<< ", frame.mean()
cv.ShowImage('Image', cv.fromarray(frame))
fp.flush()
self.frame_idx += 1
if have_mov:
self.prev_gray = np.float32(frame_gray[:])
# handle events for abort
k = cv.WaitKey (int(1000 / ap['afps']))
if k % 0x100 == 27:
# user has press the ESC key, so exit
break
del fp
if ap['display']:
cv2.destroyAllWindows()
fnfrm1='/home/ar/video/drone_project/2/frame_364_640x480.3.png'
fnfrm2='/home/ar/video/drone_project/2/frame_420_640x480.3.png'
# fnfrm1='/home/ar/data/UAV_Drone_Project/VID_20150324_012829/image-01-013.jpeg'
# fnfrm2='/home/ar/data/UAV_Drone_Project/VID_20150324_012829/image-01-014.jpeg'
xy_frm1=np.array([825, 618])
xy_frm2=np.array([1147, 675])
if __name__=='__main__':
frm1=cv2.imread(fnfrm1, 0)
frm2=cv2.imread(fnfrm2, 0)
frm1=frm1.astype(np.float)
frm2=frm2.astype(np.float)
hann=cv2.createHanningWindow((frm1.shape[1], frm1.shape[0]), cv2.CV_64F)
hann2=cv2.createHanningWindow((100,100), cv2.CV_64F)
hann2_nrm=cv2.normalize(hann2, None, 0,255, cv2.NORM_MINMAX, cv2.CV_8U)
cv2.imshow("F**K", hann2_nrm)
shift = cv2.phaseCorrelate(frm1, frm2, hann)
dxy=shift[0]
print shift
print xy_frm2-xy_frm1
hann_nrm=cv2.normalize(hann, None, 0,255, cv2.NORM_MINMAX, cv2.CV_8U)
frm1_nrm=cv2.normalize(frm1, None, 0,255, cv2.NORM_MINMAX, cv2.CV_8U)
frm2_nrm=cv2.normalize(frm2, None, 0,255, cv2.NORM_MINMAX, cv2.CV_8U)
cv2.imshow("hanning", hann_nrm)
cv2.imshow("frame #1", frm1_nrm)
# cv2.imshow("frame #2", frm2_nrm)
frm2_nrm_shift=np.roll(frm2_nrm, int(math.floor(-dxy[0])), 1)
frm2_nrm_shift=np.roll(frm2_nrm_shift, int(math.floor(-dxy[1])), 0)
img2F=np.abs(np.fft.fftshift(np.fft.fft2(img2)))
img1Fs=ph.nrmMat(np.abs(np.log(img1F+1.0)))
img2Fs=ph.nrmMat(np.abs(np.log(img2F+1.0)))
cv2.imshow("win-img-12", np.dstack((img1,img2,img1)).astype(np.uint8) )
# cv2.imshow("win-fft-12", np.dstack((img1Fs,img2Fs,img1Fs)) )
CC,dxy,ccval=ph.phaseCorr(img1, img2, isDebug=False)
print dxy, ccval
# cv2.imshow("win-CC", nrmMat(CC) )
img2Shift=np.roll(np.roll(img2,dxy[0],1), dxy[1], 0)
# cv2.imshow("win-shift-12", np.dstack((nrmMat(img1),nrmMat(img2Shift),nrmMat(img1))) )
#
# Rotate-Scale
img1FLP,MM=ph.getLogPolar(img1F.copy())
img2FLP,MM=ph.getLogPolar(img2F.copy())
# hwLP=getCylHanning(img1FLP.shape, axis=0)
hwLP=cv2.createHanningWindow(img1FLP.shape[::-1], cv2.CV_64F)
img1FLP=img1FLP*hwLP
img2FLP=img2FLP*hwLP
#
img1FLPN=ph.nrmMat( np.abs(np.log(img1FLP+1.0)) )
img2FLPN=ph.nrmMat( np.abs(np.log(img2FLP+1.0)) )
# cv2.imshow("win-Hann-LP", nrmMat(hwLP) )
cv2.imshow("win-fft-LP-img1", np.dstack((img1FLPN,img2FLPN,img1FLPN)) )
CCRS,dxyRS,ccvalRS=ph.phaseCorr(img1FLP,img2FLP)
#
dAng=(360./CCRS.shape[0])*dxyRS[1]
dScl=np.exp(float(dxyRS[0])/MM)
print dxyRS, ccvalRS, ", dAng = ", dAng, ", dScale=", dScl
#
rotm2=cv2.getRotationMatrix2D((img2.shape[1]/2, img2.shape[0]/2), -dAng, 1./dScl)
print rotm2
def load_H(self, H_file):
self.H = np.load(H_file)
h, w = self.H.shape[:2]
self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
def preprocess(self, img, size):
img = np.log(np.float32(img)+1.0)
img = (img-img.mean()) / (img.std()+eps)
win = cv2.createHanningWindow((size[0], size[1]), cv2.CV_32F)
return img*win
def _process_phase_corr(**kwargs):
if not HAVE_CV:
print "WARNING: You must install OpenCV in order to analyze or view!"
return
ap = kwargs
ap = {'fps':24, 'grid_divs_x':8, 'grid_divs_y':8, 'stride':STRIDE, 'mode':'analysis', 'display': True}
# fstring = '~/dev/git/synaesthesia/films/Synae_test_v2.mov'
fstring = '~/Movies/action/JoggersRaining/JoggersRaining.mov'
# fstring = '~/Pictures/iPhoto Library/Masters/2012/05/31/20120531-210318/IMG_0140.MOV'
movie_path = ap['movie_path'] = os.path.expanduser(fstring)
data_path = ap['data_path'] = os.path.expanduser('~/dev/git/synaesthesia/films/Synae_test_v2.pkl')
verbose = True
if (ap['movie_path'] is None) or (ap['data_path'] is None):
print "ERROR: Must supply both a movie and a data path!"
return
capture = cv2.VideoCapture(ap['movie_path'])
fps = ap['fps']
grid_x_divs = ap['grid_divs_x']
grid_y_divs = ap['grid_divs_y']
frame_width = int(capture.get(cv.CV_CAP_PROP_FRAME_WIDTH))
frame_height = int(capture.get(cv.CV_CAP_PROP_FRAME_HEIGHT))
frame_size = (frame_width, frame_height)
grid_width = int((frame_width/grid_x_divs)/DSF)
grid_height = int((frame_height/grid_y_divs)/DSF)
grid_size = (grid_width, grid_height)
centers_x = range((frame_width/16/DSF),(frame_width/DSF),(frame_width/8/DSF))
centers_y = range((frame_height/16/DSF),(frame_height/DSF),(frame_height/8/DSF))
if verbose:
print fps, ' | ', frame_size, ' | ', grid_size
# container for prev. frame's grayscale subframes
prev_sub_grays, prev_grid_xvals, prev_grid_yvals = [], [0 for i in range(64)], [0 for i in range(64)]
# last but not least, get total_frame_count and set up the memmapped file
dur_total_secs = int(capture.get(cv.CV_CAP_PROP_FRAME_COUNT) / fps)
stride_frames = ap['stride']
# if ap['duration'] < 0:
dur_secs = dur_total_secs
# else:
# dur_secs = ap['duration']
offset_secs = 0 # min(max(ap['offset'], 0), dur_total_secs)
dur_secs = min(max(dur_secs, 0), (dur_total_secs - offset_secs))
offset_strides = int(offset_secs * (fps / stride_frames))
dur_strides = int(dur_secs * (fps / stride_frames))
offset_frames = offset_strides * stride_frames
dur_frames = dur_strides * stride_frames
if verbose:
print '%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%'
print 'FRAMES: ', int(capture.get(cv.CV_CAP_PROP_FRAME_COUNT))
print 'DUR TOTAL: ', dur_total_secs
print "OFFSET (SECONDS): ", offset_secs
print "OFFSET (STRIDES): ", offset_strides
print "OFFSET (FRAMES): ", offset_frames
print "DUR (SECONDS): ", dur_secs
print 'DUR (STRIDES): ', dur_strides
print 'DUR (FRAMES): ', dur_frames
print 'XDIVS: ', grid_x_divs
print 'YDIVS: ', grid_y_divs
print "FPS: ", fps
print "stride_frames: ", stride_frames
# set up memmap
if ap['mode'] == 'playback' and ap['display'] == True:
fp = np.memmap(data_path, dtype='float32', mode='r+', shape=((offset_strides + dur_strides),64,2))
else:
fp = np.memmap(data_path, dtype='float32', mode='w+', shape=(dur_strides,64,2))
# set some drawing constants
frame_idx = offset_frames
end_frame = offset_frames + dur_frames
if ap['display']:
cv.NamedWindow('Image', cv.CV_WINDOW_AUTOSIZE)
capture.set(cv.CV_CAP_PROP_POS_FRAMES, offset_frames)
fhann = cv2.createHanningWindow((frame_width/DSF,frame_height/DSF), cv2.CV_32FC1)
ghann = cv2.createHanningWindow((grid_width/DSF,grid_height/DSF), cv2.CV_32FC1)
# load arrays with default vals
for row in range(grid_y_divs):
for col in range(grid_x_divs):
#prev_sub_grays += [np.float32(frame_gray_pyr[(row*grid_height):((row+1)*grid_height), (col*grid_width):((col+1)*grid_width)])]
# sub_grays += [np.zeros_like(ghann, dtype=np.int8)]
prev_sub_grays += [np.zeros_like(ghann, dtype=np.int8)]
xval = int(0+centers_x[col])
yval = int(0+centers_y[row])
frame_gray_pyr = np.zeros_like(fhann, dtype=np.int8)
prev_frame_gray_pyr = np.zeros_like(fhann, dtype=np.int8)
ds_grid_height = grid_height/DSF
ds_grid_width = grid_width/DSF
while frame_idx < end_frame:
if ((frame_idx % stride_frames) == 0):
t = clock()
print '1. fr. idx: ', frame_idx, ' (', frame_idx / float(end_frame), ' | ', end_frame, ')'
# grab next frame
ret, frame = capture.read()
dt = clock() - t
print 'time: %.1f ms' % (dt*1000)
t = clock()
if frame is None:
print 'Frame error! Exiting...'
break # no image captured... end the processing
prev_frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
if DSF > 1:
prev_frame_gray_pyr = cv2.pyrDown(prev_frame_gray)
else:
prev_frame_gray_pyr = prev_frame_gray #[:]
# grid stage (prev frame)
for row in range(grid_y_divs):
for col in range(grid_x_divs):
cell = ((row*grid_x_divs)+col)
prev_sub_grays[(row*grid_x_divs)+col] = prev_frame_gray_pyr[(row*ds_grid_height):((row+1)*ds_grid_height), (col*ds_grid_width):((col+1)*ds_grid_width)]
dt = clock() - t
print '2. time: %.1f ms' % (dt*1000)
if (frame_idx % stride_frames) == 1:
# grab next frame
t = clock()
ret, frame = capture.read()
if frame is None:
print 'Frame error! Exiting...'
break # no image captured... end the processing
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
if DSF > 1:
frame_gray_pyr = cv2.pyrDown(frame_gray)
else:
frame_gray_pyr = frame_gray #[:]
# grid stage (full)
for row in range(grid_y_divs):
for col in range(grid_x_divs):
if ap['mode'] == 'playback':
cell = ((row*grid_x_divs)+col)
gres = fp[(frame_idx/stride_frames)][cell]
else:
# print '--grid-----------'
# print (row*ds_grid_height)
# print (col*ds_grid_width)
# print (row,col)
sub_gray = np.float32(frame_gray_pyr[(row*ds_grid_height):((row+1)*ds_grid_height), (col*ds_grid_width):((col+1)*ds_grid_width)])
# print sub_gray.shape
# print np.float32(prev_sub_grays[(row*grid_x_divs)+col]).shape
# print ghann.shape
gres = cv2.phaseCorrelateRes(np.float32(prev_sub_grays[(row*grid_x_divs)+col]), sub_gray, ghann)
dflag = 0
if abs(gres[1]) > 0.9:
fp[(frame_idx/stride_frames)][(row*grid_x_divs)+col] = [(gres[0][0]/grid_width),(gres[0][1]/grid_height)]
dflag = 1
else:
fp[(frame_idx/stride_frames)][(row*grid_x_divs)+col] = [0,0]
if ap['display'] == True and dflag > 0:
xvec = gres[0][0]*1000.0
yvec = gres[0][1]*1000.0
#sqrt preserving signs:
if xvec < 0:
xval = int((-1*((-1*xvec) ** 0.5))+centers_x[col])
else:
xval = int((xvec ** 0.5)+centers_x[col])
if yvec < 0:
yval = int((-1*((-1*yvec) ** 0.5))+centers_y[row])
else:
yval = int((yvec ** 0.5)+centers_y[row])
# xval = int(xvec)+centers_x[col]
# yval = int(yvec)+centers_y[row]
#delta = ((xval - prev_grid_xvals[(row*grid_x_divs)+col]) ** 2.0) + ((yval - prev_grid_yvals[(row*grid_x_divs)+col]) ** 2.0)
#print delta
#if delta < MAXDIST:
cv2.line(frame_gray_pyr, (centers_x[col], centers_y[row]), (xval, yval), (255,255,255))
#prev_grid_xvals[(row*grid_x_divs)+col] = xval
#prev_grid_yvals[(row*grid_x_divs)+col] = yval
print
dt = clock() - t
print 'time: %.1f ms' % (dt*1000)
#### SHOW
if ap['display']:
print "show"
cv.ShowImage('Image', cv.fromarray(frame_gray_pyr))
fp.flush()
if ((frame_idx % stride_frames) == 0):
frame_idx += 1
elif (frame_idx % stride_frames) == 1:
frame_idx += 5
capture.set(cv.CV_CAP_PROP_POS_FRAMES, frame_idx)
# print "fr. idx.: ", frame_idx
# prev_frame_gray_pyr = np.float32(frame_gray_pyr[:])
# handle events for abort
k = cv2.waitKey (1)
if k % 0x100 == 27:
# user has press the ESC key, so exit
break
del fp
if ap['display']:
cv2.destroyAllWindows()
def preprocess(self, img, size):
img = np.log(np.float32(img)+1.0)
img = (img-img.mean()) / (img.std()+eps)
win = cv2.createHanningWindow((size[0], size[1]), cv2.CV_32F)
return img*win
import cv2
import numpy as np
img = cv2.imread('C:\\Users\\ankitdeora2856\\Desktop\\pic1.jpg')
res = cv2.getRectSubPix(img,(200,200),(320,500))
win = cv2.createHanningWindow((500, 300), cv2.CV_32F)
blur = cv2.GaussianBlur(img,(-1,-1),10)
win0 = cv2.resize(win,None,fx=0.5, fy=0.5, interpolation = cv2.INTER_CUBIC)
#cv2.namedWindow('img',cv2.WINDOW_NORMAL)
cv2.imshow('img',win0)
cv2.waitKey(0)
cv2.destroyAllWindows()
def POC(a, b):
imshowflag = 1 # show the processing image
a = a.astype(np.float32)
b = b.astype(np.float32)
height, width = a.shape
hann = cv2.createHanningWindow((height, width), cv2.CV_64F)
# rhann = np.sqrt(hann)
# Windowing and FFT
G_a = np.fft.fft2(a * hann)
G_b = np.fft.fft2(b * hann)
# 1. Get Rotation and Scaling Error
# 1.1: Frequency Whitening
LA = np.fft.fftshift(np.log(np.absolute(G_a) + 1))
LB = np.fft.fftshift(np.log(np.absolute(G_b) + 1))
# 1.2: Log polar
cx = width / 2
cy = height / 2
Mag = width / math.log(width)
LPA = cv2.logPolar(LA, (cy, cx), Mag, cv2.INTER_LINEAR)
LPB = cv2.logPolar(LB, (cy, cx), Mag, cv2.INTER_LINEAR)
# 1.3: Filtering
# --------------- Tuning parameter ----------------
lpmin_tuning = 0.5 # default 0.5
lpmax_tuning = 0.8 # default 0.8
# -------------------------------------------------
LPmin = math.floor(Mag * math.log(lpmin_tuning * width / 2.0 / math.pi))
LPmax = min(width, math.floor(Mag * math.log(width * lpmax_tuning / 2)))
assert LPmax > LPmin, 'Invalid condition!\n Enlarge lpmax tuning parameter or lpmin_tuning parameter'
Tile = np.repeat([0.0, 1.0, 0.0],
[LPmin - 1, LPmax - LPmin + 1, width - LPmax])
Mask = np.tile(Tile, [height, 1])
LPA_filt = LPA * Mask
LPB_filt = LPB * Mask
# 1.4: Phase Correlate to Get Rotation and Scaling
Diff, peak = PhaseCorrelation(LPA_filt, LPB_filt)
# Do not use "cv2.phaseCorrelate(LPA_filt,LPB_filt)" because the 2nd order fitting process is not suitable for this fucntion.
# Final output of scale and rotation
theta1 = 360 * Diff[1] / height
# deg
theta2 = theta1 + 180
# deg theta ambiguity
#invscale = math.pow(float(width),Diff[0]/float(width))
invscale = math.exp(Diff[0] / Mag)
# print('Theta? ',-theta1,'Scale ',1/invscale)
# 2.1: Correct rotation and scaling
b1 = Warp_4dof(b, 0, 0, theta1 * math.pi / 180, invscale)
b2 = Warp_4dof(b, 0, 0, theta2 * math.pi / 180, invscale)
# 2.2 : Translation estimation
diff1, peak1 = cv2.phaseCorrelate(
a, b1) #diff1, peak1 = PhaseCorrelation(a,b1)
diff2, peak2 = cv2.phaseCorrelate(
a, b2) #diff2, peak2 = PhaseCorrelation(a,b2)
# Use cv2.phaseCorrelate(a,b1) because it is much faster
# 2.3: Compare peaks and choose true rotational error
if peak1 > peak2:
Trans = diff1
peak = peak1
theta = -theta1
else:
Trans = diff2
peak = peak2
theta = -theta2
if theta > 180:
theta -= 360
elif theta < -180:
theta += 360
if 0:
# Imshow
plt.subplot(5, 2, 1)
plt.imshow(LA, vmin=LA.min(), vmax=LA.max())
plt.subplot(5, 2, 2)
plt.imshow(LB, vmin=LB.min(), vmax=LB.max())
plt.subplot(5, 2, 3)
plt.imshow(LPA, vmin=LPA.min(), vmax=LPA.max(), cmap="gray")
plt.subplot(5, 2, 4)
plt.imshow(LPB, vmin=LPB.min(), vmax=LPB.max(), cmap="gray")
plt.subplot(5, 2, 5)
plt.imshow(LPA_filt,
vmin=LPA_filt.min(),
vmax=LPA_filt.max(),
cmap="gray")
plt.subplot(5, 2, 6)
plt.imshow(LPB_filt,
vmin=LPB_filt.min(),
vmax=LPB_filt.max(),
cmap="gray")
plt.subplot(5, 2, 7)
plt.imshow(b1, vmin=b1.min(), vmax=b1.max(), cmap="gray")
plt.subplot(5, 2, 8)
plt.imshow(b2, vmin=b2.min(), vmax=b2.max(), cmap="gray")
return [Trans[0], Trans[1], theta, 1 / invscale], peak
img2F = np.abs(np.fft.fftshift(np.fft.fft2(img2)))
img1Fs = ph.nrmMat(np.abs(np.log(img1F + 1.0)))
img2Fs = ph.nrmMat(np.abs(np.log(img2F + 1.0)))
cv2.imshow("win-img-12", np.dstack((img1, img2, img1)).astype(np.uint8))
# cv2.imshow("win-fft-12", np.dstack((img1Fs,img2Fs,img1Fs)) )
CC, dxy, ccval = ph.phaseCorr(img1, img2, isDebug=False)
print dxy, ccval
# cv2.imshow("win-CC", nrmMat(CC) )
img2Shift = np.roll(np.roll(img2, dxy[0], 1), dxy[1], 0)
# cv2.imshow("win-shift-12", np.dstack((nrmMat(img1),nrmMat(img2Shift),nrmMat(img1))) )
#
# Rotate-Scale
img1FLP, MM = ph.getLogPolar(img1F.copy())
img2FLP, MM = ph.getLogPolar(img2F.copy())
# hwLP=getCylHanning(img1FLP.shape, axis=0)
hwLP = cv2.createHanningWindow(img1FLP.shape[::-1], cv2.CV_64F)
img1FLP = img1FLP * hwLP
img2FLP = img2FLP * hwLP
#
img1FLPN = ph.nrmMat(np.abs(np.log(img1FLP + 1.0)))
img2FLPN = ph.nrmMat(np.abs(np.log(img2FLP + 1.0)))
# cv2.imshow("win-Hann-LP", nrmMat(hwLP) )
cv2.imshow("win-fft-LP-img1", np.dstack((img1FLPN, img2FLPN, img1FLPN)))
CCRS, dxyRS, ccvalRS = ph.phaseCorr(img1FLP, img2FLP)
#
dAng = (360. / CCRS.shape[0]) * dxyRS[1]
dScl = np.exp(float(dxyRS[0]) / MM)
print dxyRS, ccvalRS, ", dAng = ", dAng, ", dScale=", dScl
#
rotm2 = cv2.getRotationMatrix2D((img2.shape[1] / 2, img2.shape[0] / 2),
-dAng, 1. / dScl)
def getLittleF(gray):
inter = np.log(gray.astype('float64') + 1)
inter2 = (inter - np.amin(inter)) / np.amax(inter)
return cv2.createHanningWindow((inter2.shape[1], inter2.shape[0]), cv2.CV_64F) * inter2
def init_hanning_window(self):
self.win = cv2.createHanningWindow(tuple(self.size), cv2.CV_32F)
return self
